A short piece of fiction, written by API Evangelist. The reader is imagined; the post, and every number and quotation from it, is real: "What To Check Before You Call Your API Production Ready," Kin Lane, November 4, 2026, https://apievangelist.com/2026/11/04/what-to-check-before-you-call-your-api-production-ready/
Marisol read the post on a Wednesday afternoon, between a failing deploy and a standup she was already late for. She was the only backend engineer at a team of five, and her API had been called production ready for about a month, mostly because nobody had a better word for it.
The part that stopped her was the idempotency paragraph. She knew the idea from payments, and had never thought it applied to her own endpoints. Her service created reports. If a client called twice, it made two reports. The post said "agents are going to knock on your door one, two, three, four times", and she pictured four identical reports sitting in a customer's account, each one costing compute she paid for.
She did not enjoy the next part. The post compared Sinch and Vonage, two providers at 71.5 and 71.3 on the Kin Score, and traced nine of the fifteen points between their Agent Readiness scores of 56 and 41 to a single characteristic: 9 of 9 for idempotency on one side, 0 of 9 on the other. Marisol felt a flicker of irritation. A number like that looks like a verdict, and she had seen what verdicts do to a roadmap. Someone would paste a score into a planning doc and it would become a goal.
Then she read the author's own caveat, that a zero reflects what the scorer could find published and that "the fix might be mostly documentation." That calmed her a little, and it also made her uneasy in a different way. If the fix might be mostly documentation, then her own gaps might be mostly silence. Her API did a few sensible things that she had never written down anywhere a stranger could find them.
She opened her error handler and looked at what clients got when a report job failed. It was a bare 500 with a string. The post had a line for that, saying an agent cannot read between the lines of "a vague 500" the way a developer can. She had read between those lines herself a dozen times, with logs open and a coffee. An agent would have neither.
She did not agree with all of it. The post could not know that her biggest worry was a customer with forty thousand rows, not a polite agent retrying. She said so out loud to the empty meeting room. And the post said so too, in a way: "None of this tells you that you are safe." It was oddly reassuring to be told that by the thing selling her a checklist.
She was not going to adopt Problem Details this week, or publish limits, or add her API to a catalog to see what the reading said. Those were real items and she copied all of them into a note for later.
But one item on the startup checklist was small enough to do before standup. Decide what happens on a retried write, and write it down. She opened the README for the report endpoint and added a paragraph under a new heading, plainly stating that today a repeated call creates a second report, and that this is a known gap. Writing it was uncomfortable. She committed it anyway, then went to find out what the team wanted to do about it.